节点文献
语音信息隐藏与分析技术研究
Research on Audio Steganography and Steganalysis Techniques
【作者】 戚银城;
【导师】 苑津莎;
【作者基本信息】 华北电力大学(河北) , 电工理论与新技术, 2009, 博士
【摘要】 信息隐藏作为保障信息安全的新技术,吸引了国内外众多学者的关注,已成为信息安全领域的研究热点。信息隐藏是要隐藏信息的存在性,信息隐藏分析是对信息隐藏系统进行攻击的技术。论文围绕如何提高隐藏容量、透明性、鲁棒性和如何提高信息隐藏分析的准确性,对以语音为载体的信息隐藏与隐藏分析技术展开研究。针对回声信息隐藏方法隐藏容量小的不足,提出一种回声多进制信息隐藏系统。仿真研究了延时时间、分段长度、回声的衰减系数等关键参数对隐藏信息恢复率的影响。通过引入32种延时的前向-后向回声实现了32进制信息隐藏,在透明性较好的情况下,使隐藏信息的容量较单回声方法提高了5倍,恢复率达到99%。对隐藏系统的健壮性进行了仿真分析,如抗噪声、再抽样、滤波、ADPCM压缩等,具有较好的效果。提出了一种由语音质量评价和心理声学模型联合控制嵌入强度的语音隐藏方法,克服了对透明性进行评价时完全依靠主观试听的缺点。给出了由语音质量评价和心理声学模型联合控制嵌入强度的算法框架,并详细描述了该算法从秘密信息的嵌入到提取的全过程。最后从信息隐藏的嵌入容量、隐蔽性、鲁棒性三个方面验证了本文算法的性能,同时与采用单一嵌入强度的叠加法进行了比较。实验结果表明本文算法在隐藏容量、隐蔽性及鲁棒性方面均有较好的表现,且性能优于采用单一嵌入强度的叠加算法。针对常用的小波域语音信息隐藏方法,研究了对应的隐藏分析技术,利用隐藏信息前后语音载体统计特征的变化来区分语音信号中是否有隐藏信息。针对小波域加性噪声嵌入模型,提出两种有效的分类特征:语音信号小波子带系数直方图联合特征和小波域幅度共生矩阵特征。对于训练和测试语音集,完成了特征的提取并采用BP网络进行分类,实验结果验证了两种特征的有效性。对于小波域乘性嵌入模型,为了克服高阶矩特征、小波子带系数直方图联合特征和小波域幅度共生矩阵特征分类正确率低的不足,提出采用同态处理的方法,得到较好的分类效果。由于语音信号小波子带系数直方图联合特征和小波域幅度共生矩阵特征的维数较高,不利于分类,本文提出了一种基于特征选择技术的小波域语音隐藏分析系统模型。提取待分析语音信号的特征后,对特征进行主成分分析处理或因子分析处理,以降低特征的维数,再使用支持向量机作为分类器来对待分析语音信号进行分类检测。实验结果表明,对3种小波域常用的语音信息隐藏算法的检测正确率均大于95%。
【Abstract】 Information Hiding (steganography), a new technology in information security fields, interests many researchers in the world. The goal of steganography is to hide the existence of information, steganalysis is a technology to attack the information hiding system and detect the hiding information. Audio steganography and steganalysis technology are studied in this paper. On the basis of the analysis of main theoretical problem and the research of some technology problems, some algorithms are proposed to improve the information hiding capacity, transparency, robustness and the accuracy of wavelet domain speech steganalysis.An echo multi-ary data hiding system is proposed to overcome the disadvantages of low hiding capacity of echo hiding system. The influence of key parameters of the system to the restoration rate of information is studied, such as delay time, segment length, and decay amplitude. By introducing forward-backward echo kernel with 32 time delays, a 32-ary information hiding system is realized and the hiding capacity is five times of single echo hiding system. When the sample rate of cover-audio is 8 kHz, restoration rate of information reaches to 99%. The robustness of the steganography system is simulated too, such as attacks of white noise, resampling (upsampling, down sampling), filtering (low pass filtering, high pass filtering), ADPCM compressing.An adaptive speech information hiding method, in which its embedding amplitude is controlled by audio quality assessments and psychology model, is proposed. This method can overcome the disadvantages of subjective listening in judging the transparency of stego-audio. The procedure of the algorithm from information embedding to information detecting is dressed in detail. At last, the embedding capacity, transparency, robustness performances are studied and compared with the constant amplitude embedding method. Simulation results show that this algorithm is effective.Audio steganalysis technologies for wavelet domain embedding methods are studied, the difference of characteristic features after information embedded is utilized to determine if information is embedded to the audio signal. Two effective classification features are proposed for additive noise embedding model in wavelet domain. They are speech wavelet subband coefficients histogram union features and wavelet domain amplitude co-occurrence matrix features. These two features are extracted and used to classify the audio signal by BP neural networks. Simulation results show the effectiveness of these two features. For multiplicative noise embedding model in wavelet domain, high order statistical moments features, wavelet subband coefficients histogram union features and wavelet domain amplitude co-occurrence matrix features have low accuracy in classification. Homomorphic processing is used to overcome the problem of low steganalysis accuracy. The test audio signal is firstly calculated its absolute value and logarithm. Multiplicative noise is changed to additive noise and the classification accuracy is improved.A wavelet domain audio steganalysis system based on feature selection technology is constructed. Classification features of audio signals to be analyzed are extracted. And then these features are processed by the principle component analysis or the factor analysis methods. The dimensions of these features are reduced evidently. Support vector machine is used as the classifier to classify the cover-audio and the stego-audio. Simulation results show that the detection rate reaches to 95% for three wavelet domain steganography methods.
【Key words】 Information hiding; Steganalysis; Echo hiding; Histogram statistical moments; Amplitude co-occurrence matrix;
- 【网络出版投稿人】 华北电力大学(河北) 【网络出版年期】2009年 11期
- 【分类号】TP309.7
- 【被引频次】4
- 【下载频次】402
- 攻读期成果